Finding Spin Glass Ground States Through Deep Reinforcement Learning and Graph Neural Networks

Tomasz Śmierzchalski

supervisor: Tomasz Trzciński

Spin glasses are disordered magnets with random interactions that are, generally, in conflict with each other. Finding the ground state of a spin glass is not only an important problem in modern physic but is also connected to a wide array of hard optimization problems (such as travelling salesman). It is still an unsolved problem, but there are many heuristic algorithms useful in finding an approximate solution.


One of such is recently developed DIRAC - a deep reinforcement learning framework, which can be trained purely on small-scale spin glass instances and then applied

to arbitrarily large ones. Connected with it is SGNN (Spin Glass Neural Network), a graph neural network used to encode spin glass instances. This poster will be focused on describing SGNN and DIRACT, theirs scalability and potential shortcomings